OpenAI Launches ChatGPT Images 2.5 With Sketch-Based Creation and Local Image Editing

OpenAI introduced ChatGPT Images 2.5 on September 8, 2026, positioning the update around more personalized image creation from ideas, sketches, and reference photos. The launch centers on helping users guide image generation with visual input rather than text alone.

The announcement’s public details remain focused on the product experience. OpenAI’s presentation emphasizes polished results that better reflect a user’s original concept, while independent testing provides the clearest description of how the workflow operates. The update arrives as image generators compete on control, consistency, and ease of editing, rather than image creation alone.

ChatGPT Images 2.5 adds Sketch, an in-chat drawing tool that converts rough marks into a more detailed generated image. Users activate it by typing @Sketch, draw inside a pop-up window, and describe the desired result with a prompt.

A hands-on test turned a simple computer-mouse drawing of a cat into a realistic photograph after the tester supplied additional instructions. The feature therefore treats a crude sketch as a visual layout or subject cue, giving users a faster starting point than describing every shape in words. The product also addresses a familiar barrier in image generation: users can communicate composition visually even when they can't draw well.

ChatGPT Images 2.5 also lets users comment directly on parts of an image they want changed. In testing, a request attached to the cat's eyes changed their color from blue to green, showing how local instructions can replace repeated full-image prompts.

That editing model matters for practical work such as illustrations, concept art, social graphics, and product mockups, where small revisions often consume more time than the first generation. OpenAI's examples highlight images shaped by sketches and reference material, while the hands-on account points to localized feedback as the more immediate usability improvement. The combination gives users two forms of control: a rough visual plan before generation and precise corrections afterward.

The consumer-facing launch confirms ChatGPT Images 2.5 as a product feature, but developer-facing model details remain unsettled. Earlier reporting found no matching GPT Image 2.5 model ID in OpenAI's SDKs, changelog, or documentation, while the GPT Image 2 API remained the documented baseline before the announcement.

Morphic's review of anonymous August testing referenced possible checkpoints, output limits of up to 2K, and generation of as many as 8 images with shared characters and style. Those details were explicitly unconfirmed by OpenAI. WaveSpeed's earlier guidance was to keep model IDs configurable and build against GPT Image 2 rather than depend on an unannounced 2.5 endpoint. OpenAI's product announcement changes the status of the feature, but it doesn't by itself establish API availability or technical specifications.

ChatGPT Images 2.5 shifts image generation toward an iterative design conversation. A user can sketch a rough arrangement, request a visual style, inspect the result, and mark a specific area for revision. That workflow brings image tools closer to familiar design software, where users provide structure and refine selected elements instead of restarting from scratch.

The feature also broadens access to visual prototyping. A person with limited drawing ability can still indicate pose, placement, or object relationships, then let the model handle rendering. The practical value will depend on how consistently the system preserves requested details across edits, a question the launch examples begin to address but do not fully resolve.

The immediate test for ChatGPT Images 2.5 is adoption of Sketch and localized editing in everyday image tasks. Users will judge the update by whether rough input survives translation into a useful composition and whether small corrections avoid unwanted changes elsewhere.

Developers face a separate decision around access. Until OpenAI publishes model identifiers, API documentation, pricing, and limits, production integrations should keep image-model configuration flexible. The launch establishes a new ChatGPT image experience, while questions about a standalone GPT Image 2.5 API, technical performance, and rollout remain open.